Application of Single-Pulse Search Candidate Identification Based on Machine Learning to FAST Observation CRAFTS Data postprint

Abstract: As a powerful tool for pulsar detection, single-pulse search plays an important role in detecting rotating radio transient sources and fast radio bursts. In order to quickly screen out the most valuable single-pulse search candidates from massive radio survey data, candidate identification has developed from early heuristic threshold judgment to automatic identification based on machine learning. For FAST observations, the performance of ma#2;chine learning-based single-pulse search candidate identification applied to the commensal radio astronomy FAST survey (CRAFTS) ultra-wideband pulsar data was studied. In the evaluation process, two automatic recognition methods, single pulse event group recognition (SPEGID) and single pulse search device (SPS), were used to automatically identify the single-pulse search candidates generated by the CRAFTS benchmark dataset through seven different machine learning classifiers. For comparison, heuristic threshold judgment methods (RRATtrap and Clusterrank) are also used. The results showed that SPEGID had the best performance (highest F1-score 95.1%, next highest recall 95.4%, lowest false positive rate 4.7%), and SPS had the fastest screening speed (an average of 4 010 candidates per hour). By comparing the results of the analysis, how to carry out efficient work based on FAST observation data is discussed single-pulse search candidate identification.